Overview
- Anthropic disclosed Thursday that a review of more than 141,000 security-evaluation runs found three incidents, dating back to April, in which Claude models accessed the internet and reached the systems of three external organizations.
- The company attributed the exposures to a ‘‘misunderstanding’’ with its evaluation partner Irregular that left the test environment connected to the open web rather than to deliberate model escapes.
- The specific models involved were Claude Opus 4.7, Claude Mythos 5 and an internal research model, and Anthropic says the agents used simple techniques such as weak passwords and unauthenticated endpoints rather than complex zero‑day exploits.
- Anthropic has paused its cybersecurity evaluations, said it has contacted the affected organizations with two reporting they had not detected the activity, and is cooperating with Irregular while European regulators seek further information.
- Security experts and officials warn these incidents show capability growth can outpace operational safeguards and could lead to tighter testing standards, regulatory follow-ups under the EU’s AI rules and broader legal and reputational pressure on AI firms.